OpenAdapt
1,648 stars on OpenAdapt, 258 forksRun locally

Verified last-mile execution for agents.

Give agents verified hands for the GUI your APIs can't reach.

Compile a demonstration into a program an agent can invoke. Healthy runs make no model calls. Uncertainty escalates. Humans audit. You can run that loop today.

Run locallyView source

This page is the free MIT runtime for calling agents. OEM, RCM, BPO, and services go to Partners.

Try the halt, then MCP

openadapt quickstart --break-it
claude mcp add openadapt -- \ uvx --from 'openadapt-agent[tutorial]' openadapt-agent \ serve --allow-run

Full install on /start. Outcome contract in agents.txt.

No generative-model API calls on the default healthy replayWatch the 55s storyMIT-licensed runtime

A 55-second story

She shows the task once. The program checks the record.

Maria shows the task once. It watches and turns it into a small program that runs with no AI in the loop. Before it says anything worked, it checks the record, because the screen can lie. An expired policy stops the workflow. She decides on her phone.

Reenactment. Product footage uses synthetic patient data. The Maria scenes and music use generative AI.

OpenAdapt is free to use locally. pip install openadapt, then openadapt quickstart.

Start freeView source
Read the transcript
  1. Every shift, Maria retypes the same patient details into the same five screens.
  2. She'll do it about a hundred times today.
  3. None of it is hard, but she's tired, and one wrong click files the wrong note in the wrong chart.
  4. So one afternoon, she shows OpenAdapt the task instead.
  5. It watches and turns it into a small program.
  6. Now the program runs on its own, no AI in the loop.
  7. And before it says anything worked, it checks the patient's record.
  8. When OpenAdapt sees that the policy has expired, it stops before it sends anything.
  9. The program asks Maria to decide.
  10. She checks the chart on her phone and approves.
  11. The decision is on the record, and Maria gets on with her day.
  12. OpenAdapt is free to use locally.
  13. Your first workflow takes about five minutes.
  14. Get started at openadapt.ai/start.

See it run

From a demonstrated task to a verified result

The same program, twice: once proving the change landed, once refusing a result its independent check contradicts.

In a fault-injection study, screen-only checking accepted 54 of 90 wrong results. A complete system-of-record read accepted 0 of 90. Read the study

OpenAdaptObservingStep 1 of 6Observing the application
0:00 / 0:00Evidence

Guided view synchronized to the exact retained runtime timeline; raw footage remains unchanged.

  1. 1
    DemonstrateCapture the task and its evidence.
  2. 2
    ExecuteReplay the compiled workflow locally.
  3. 3
    Verify or haltProve the effect, or stop for review.
Independent checkReads the record
When that check disagreesHalts
Healthy-replay model calls0
Start locallyExplore the full demo

Same halt on your machine: openadapt quickstart --break-it. MCP for Claude Code lives in agents.txt.

How verification works

How it works

Author once. Agents run it. Humans decide only on halt.

You can run this loop today. A human authors the program. An agent invokes it. An independent check decides VERIFIED or halt.

  1. 1

    Author

    A named human demonstrates one bounded workflow. That person is authority, not the daily operator. OpenAdapt keeps the actions and the evidence around each step.

  2. 2

    Compile

    The recording becomes a program an agent can invoke. Review the steps, identities, expected results, and environment before a version is reusable.

  3. 3

    Agent invokes

    The calling agent binds parameters and runs the program. Computer-use agents are the user of OpenAdapt. They are not the executor inside it.

  4. 4

    VERIFIED or halt

    The run reports VERIFIED only when an independent check passes. Identity, effect, and judgment halts page a human. Missing parameters may come from another tool.

Open source

An agent can call it today. So can the CLI.

pip install openadapt leaves an agent-callable surface, not only a human CLI. Point Claude Code or Cursor at it with the MCP command below. Then openadapt quickstart --break-it. The default healthy replay makes no generative-model API calls. Never summarize halt as success.

claude mcp add openadapt -- \
  uvx --from 'openadapt-agent[tutorial]' openadapt-agent \
  serve --allow-run

Install OpenAdapt

One command installs the launcher, compiler, governed runtime, and browser driver used by the local tutorial. No account, API key, or hosted service is required.

Install
$pip install openadapt
Installs the OpenAdapt launcher and compiler for Python 3.10+ on Windows, macOS, or Linux. Run the engine under openadapt flow. For isolated command-line installs, the docs also cover uv tool. Read docs

Get a verified result

The bundled tutorial records, compiles, certifies, and runs a synthetic task. A separate read-only API verifies the saved result before OpenAdapt reports VERIFIED. The first browser action downloads matching Chromium once.

openadapt quickstartComplete the local lifecycle and write the bundle, report, and privacy-safe receipt.

The default healthy tutorial run makes no generative-model API calls and does not contact OpenAdapt Cloud. The report shows what ran, what OpenAdapt verified, and why it halted. The bundled workflow is a tutorial. Continue in the five-minute walkthrough. Compiler source and measured limits are in the openadapt-flow engine.

Inside the program

See the program compare recorded evidence with the live application.

OpenAdapt ties each compiled action to the evidence that created it. Play the exact trace below, or switch to the safety halt. It won’t act when the current evidence can’t prove one safe target.

Inspect the full method
Program State ConsoleRepresentative synthetic program
v38d9c…42afpublic synthetic
Program artifact
  1. Recorded evidence
  2. Compiled program v3
  3. Repair candidate v4
  4. Review and trials
  5. Active v4
One live run
  1. 1Current observation
  2. 2Resolve target
  3. 3Identity gate
  4. 4Act once
  5. 5Screen check
  6. 6Effect check
  7. RUNNING
01 / 09
ObserveInput contract acceptedevent 01

The field moved. The retained evidence still proves one target, so the run continues.

1 of 9: Input contract accepted
Exact synthetic traceTarget geometry never interpolatesNo model call in this run

A strong first workflow

Start where an error has a cost and the result can be checked

  • One person can demonstrate it from start to finish.
  • A wrong action costs money, time, or compliance standing.
  • The result can be checked independently, or by reacquiring the record.

Record a first workflow on your machine

Browser, desktop, and remote

The same loop on browser, desktop, and remote

Each workflow bundle binds one target and one execution backend. You can record, compile, and prove a result on these surfaces today. A task that crosses a boundary uses separate bundles, external orchestration, and an end-to-end result check.

  • Browser

    Any web application.

  • Native desktop

    Windows, macOS, and Linux applications.

  • Remote applications

    RDP, Citrix, and VDI sessions.

Run it on your own machine, or entirely inside your own boundary. Start locally

Production is a signed admission you earn. The public lifecycle record currently lists all seven product targets with no admission. We'll build Flow next. See the ledgerproduction-lifecycle.json

What each surface uses, and what it supports todayReview the security boundary

Data and control

Keep sensitive work in the boundary you approve

  • Customer-controlled execution for sensitive and remote work
  • Local processing for live screenshots, OCR, identities, and secrets
  • Explicit network and model policy for Standard and Regulated profiles
  • Signed, sealed, versioned, and independently verifiable artifacts
  • Evidence-backed VERIFIED or HALTED outcomes
  • MIT-licensed local runtime you can inspect and run without an account

Open the trust center